Professional Documents
Culture Documents
SPSS Advance
SPSS Advance
2011 ()
Survival analysis 37
SPSS () 73
/
2011 ()
()
-1-
2011 ()
Analysis
A l i ffor
Repeated measures data
( )
Contents
1. .
2. ?
3. Repeated measures ANOVA & Mixed model
4
4. Generalized estimating equation(GEE)
-3-
/
1
1.
Ex)
(2 )
Paired t-test, or
Wilcoxons signed rank test
Design
2.
(Repeated measures data?)
: , , , .
: , , .
-4-
2011 ()
->
, (
)
(the measures within subjects are correlated)
Example1
(55 , 55 )
1, 3, 6, 9, 12
-5-
/
1,
3, 6, 9, 12
( )
:
- (55, >55) , .
(Within subjects effect) (time effect) Main
effect
( )
- , (55 , )
.
(Between subjects
j effect) (age
g ggroup
p effect)
- (55 ,
) .
(Interaction
(I t ti effect(
ff t( ) )
-6-
2011 ()
(Interaction)
A
B
( A B
, A B
)
->
Ex)
(55 vs )
,
Blue line : 55
Red
d li
line : 55
3 6)
36)
10
-7-
/
time age(55, >55) age*time
11
example
p
12
-8-
2011 ()
3 &
3
<-
outcome
13
Example1
(55 , ) 1,
3, 6, 9, 12
( ) 5
14
-9-
/
Independent variables
: age(55 , ), time()
15
16
- 10 -
2011 ()
- , .
(Within subjects effect) (time effect)
Main
effect
- , (55 , ) ( )
.
(Between subjects effect) (age group effect)
- (55 ,
) . (Interaction effect( ) )
17
Method
h d:
18
- 11 -
/
19
: OK
R
Repeated
t d measures ANOVA
th
then, check
h k
- ,
Greenhouse-Geisser
20
- 12 -
2011 ()
(Sphericity)
(S h i it )
:
(compound symmetry)
(equal
l variance)
i
Mauchly's Test of Sphericity
21
: O
Repeated
Greenhouse-
measures
Geisser
ANOVA
: X
22
- 13 -
/
test
23
->
>
- ,
(P-value = 0.021)
- ) :
(55 , ) .
(P-value = 0.506) 24
- 14 -
2011 ()
- , (55 , )
(P-value=
. (P value 0
0.048)
048)
25
P-
P < 00.0
0 ; ,
26
- 15 -
/
- Pillai (trace)
- Wilks (lambda)
- g
Hotelling
-
SPSS default
27
.
->
>
Greenhouse-Geisser, Huyng-Feldt
28
- 16 -
2011 ()
Independent variables
: age(55 , ), time()
29
Missing data
1 21 21 9
21.9 21 9
21.9 21 9
21.9
2 21 21.9 21.9 43.8
3 20 20.8 20.8 64.6
30
- 17 -
/
b
F
Pillai
590
.590 950a
3.950
3 4 000
4.000 11 000
11.000 .032
032
Wilks .410 3.950a 4.000 11.000 .032
1
Hotelling
1 436
1.436 950a
3.950
3 4 000
4.000 11 000
11.000 .032
032
Roy 1.436 3.950a 4.000 11.000 .032
Pillai
.284 1.090a 4.000 11.000 .408
Wilks .716 1.090a 4.000 11.000 .408
1 * age_
1 Hotelling
.396 1.090a 4.000 11.000 .408
Roy .396 1.090a 4.000 11.000 .408
31
32
- 18 -
2011 ()
Example2
(more complicated data with missing)
Randomized experiment
p in which 50 male rats are
randomized to
- Control (15 rats) , Low does (18 rats),
High does (17 rats)
Treatment starts at the age
g of 45 days
y
Measurements
: everyy 10 days
y ; from dayy 50
33
Reponses
: distances between well defined points
34
- 19 -
/
Data
35
Missing (56%)
36
- 20 -
2011 ()
Rat profiles
Rats
age(days)
g ( y )
ppattern
-> much variability among rats
- various intercept
- various
i slope
l ffor age
37
: Age(days) distance
(control, low dose, high dose)
38
- 21 -
/
Repeated measures ANOVA ?
->
Why?
- Missing due to dropout
- No consideration of variability among rats
- No consideration of various correlation structure
among time (age; days)
39
40
- 22 -
2011 ()
Cf)
f Repeatedd measures ANOVA
41
Correlation structure
Correlation structure
Specifies how the observations within a
subject are related
Type
yp
Unstructured
Compound symmetry
Autoregressive AR(1)
42
- 23 -
/
Conpound symmetry( )
Any responses within the same subject has the
same correlation
Simple
p (1( parameter
p to estimate))
1
1
1
1
43
Autoregressive AR(1) (1 )
Correlation between responses depends on the
i t
interval
l off ti
time b
between
t responses
Farther apart responses => weaker correlation
Only
O l 1 parameter to estimate!
i !
1
1
2 1
3
2 1
44
- 24 -
2011 ()
Unstructured ()
All correlation
l ti coefficients
ffi i t ffree tto ttake
k any
value
E.g.,
Eg
1
0.3 1
0.1 0.5 1
0
0.05
05 0
0.2
2 0
0.4
4 1
45
()
Mixed model
46
- 25 -
/
Mixed model
Mixed
d model
d l
- (fixed effect) (random effect)
(fixed effect)
- level
( level sampling
(,
level
)
(ex) age(55 , ), , , ,
47
(random effect)
level sampling
level fixed
(, level sampling
)
48
- 26 -
2011 ()
(ex2)
21 1, 3, 6, 9, 12
21
,
.
49
Example
l Analysis
l using mixed
d model
d l
50
- 27 -
/
In Stress Example1
p ((with missing)
g)
, ,
3 6,9,
3, 6 9
51
Patients profiles
100
80
60
40
20
0
1 2 3 4 5
Patient,
P i pattern
: due to Much variability between patients
Missing data
Various Correlation structure due to Repeated
p
measurements scheduled per patient
52
- 28 -
2011 ()
(55 , ) .
(Interaction effect( ) )
53
54
- 29 -
/
55
age , stress
?
1 Reference
R f category
(1 3, 6, 9,
12 stress )
56
- 30 -
2011 ()
57
: Bonferroni
Bonferroni correction
When
h Multiple
l l testing.
HOW?
P-value * (# of multiple testings)
58
- 31 -
/
4 Generalized estimating
4.
equation(GEE)
q ( )
59
outcome
outcome type
yp
- binary
- ordinal
- nominal
60
- 32 -
2011 ()
Ex2) 1,
1 3,
3
6, 9, 12
Outcome type :
Outcome variable
: 1, 3, 6, 9, 12 (binary
data)
Independent variable
: (55 , )
61
62
- 33 -
/
:
(55 , ) .
(interaction effect() )
,
. ( main effect() )
,
.
( main effect()
( ) )
63
Analysis
l using Generalized
l d estimating equation
(GEE)
64
- 34 -
2011 ()
65
time (5 )
; P-value
l = 0.849
: age , stress
66
- 35 -
/
Summary
1. Continuous outcome
2
Repeated measures ANOVA
case
(
(imputation)
p ) impute
p
Mixed model
. 67
2. Categoricall outcome
2
GEE(Generalized
GEE(G li d estimating
ti ti equation)
ti )
GLMM(G
GLMM(Generalized
li d Li
Linear mixed
i d model)
d l)
random effect fixed effect
68
- 36 -
/
2011 ()
()
Survival analysis
( )
- 37 -
2011 ()
Survival analysis
( )
Contents
1. Survivall analysis
l ?
2. Survival data
3.
4.
5.
6. Event
7. Time dependent Cox regression model
8. More than one event
- 39 -
/
1 Survival analysis
1.
Data: Incomplete
l d
data
Incomplete?
-
<- censored data
Censoring : dont
d k
know survivall time exactly
l
Why censor ?
- Study ends
- Lost f/u
- Withdrawals
- 40 -
2011 ()
Time 0
- 41 -
/
- 42 -
2011 ()
Ti
Time 0 STUDY END
9
wGGSGGGGTSGGGGG
GGGGG
- 43 -
/
<-
Incomplete data
11
Incomplete Data
12
- 44 -
2011 ()
(, incomplete data)
13
( ) MACE(major
(ex) MACE( j adverse
d cardiac
di events)
t ) after
ft IIsolated
l t d
Aortic Valve Replacement in Patients with severe
Aortic Stenosis (A,
(A BB, C
C, D)
A (5mon)
B (24mon)
C (12mon)
D (
(9mon)
)
(: MACE, : non-MACE)
14
- 45 -
/
Subject MACE
A 5 MACE
at 5mon
B 24 Non-MACE
C 12 MACE
att 12mon
12
D 10 Non-MACE
15
Subject MACE 6
MACE
A 5 MACE MACE
at 5mon
B 24 Non-MACE Non-MACE
C 12 MACE Non-MACE
att 12mon
12
D 10 Non-MACE Non-MACE
MACEC rate
ate
at 5Mon
=
Complete
d t
data
16
- 46 -
2011 ()
Subject MACE 6 12
MACE MACE
A 5 MACE MACE MACE
at 5mon
B 24 Non-MACE Non-MACE Non-MACE
C rate
MACE ate MACE
C rate
ate
At 6Mon at 12Mon
= =?
Complete Incomplete
d t
data d t
data
17
C rate
MACE ate MACE
C rate
ate MACE
C free
ee ttime
e
At 6Mon at 12Mon =?
= =?
Complete Incomplete Incomplete
d t
data d t
data d t
data
Subject
j B: censored at 24 mon after replacement
p
Subject D: censored at 9 mon after replacement
18
- 47 -
/
Survival analysis
19
2 Data
2.
(in Survival analysis)
y
Time to event
Event
(event ; (ex) event =1, censored=0)
20
- 48 -
2011 ()
E
Event
t () Ti
Time to
t eventt
Death
D h after
f O
Op
Recurrence after Tx -
Response after Tx -
Progression after first tx
-
21
Data
Ex) MACE(major adverse cardiac events) after Isolated Aortic Valve
Replacement in Patients with severe Aortic Stenosis (A, B, C, D)
22
- 49 -
/
1 2008. 1. 21 2008. 6. 20
(MACE )
23
1 2008. 1. 21 2008. 6. 20 1 5
(MACE )
2 2007 12.
2007. 12 4 2009. 12.
2009 12 3 0 24
( F/U date)
3 2008. 3. 26 2009. 3.25 1 12
(MACE )
4 2008. 8. 31 2009. 05. 30 0 9
( F/U
/ date))
24
- 50 -
2011 ()
3
3.
- Kaplan-Meier
- Event
25
- 51 -
/
4
4.
Log-rank
k test
-
( )
(: MACE . )
Breslow test
- Event
27
cross
-
Time dependent Cox regression ,
stratified Cox regression
28
- 52 -
2011 ()
E ) Multi-vessel
Ex) M li l di
disease
non-ST-elevation
ST l i myocardial
di l
infarction single vessel multi
vessel MACE free curve
Variable :
G
Group (M
(Multi
l i vessel=1,
l 1 single
i l vessel=2)
l 2)
Time to MACE(month)
Status(MACE ; MACE=1,
MACE=1 censored=0)
29
30
- 53 -
/
31
5M MACE free
5Mo f rate ?
..
Group=1
: 5Mo MACE free rate = 92%
0.922
0.906
..
time=5.0
time=4 767
time=4.767 time=5 057
time=5.057
32
- 54 -
2011 ()
33
Group
p MACE free curve
-> Log-rank test
Results
P-value = <.001 ;
Group MACE free curve
34
- 55 -
/
5
5.
: t .
t S1(t) S2(t)
,
|Z|
p value ( SE1 SE2 S1(t) S2(t) )
p-value
35
E ) multi-vessel
Ex) li l di
disease
non ST elevation
l i myocardial
di l
infarction single vessel multi
vessel 5 MACE free rate
Variable :
G
Group (M
(Multi
l i vessel=1,
l 1 single
i l vessel=2)
l 2)
Time to MACE(month)
Status(MACE ; MACE=1,
MACE=1 censored=0)
36
- 56 -
2011 ()
37
~~~
~~~
~~~ Group=2
G 2 ((single
i l vessel)
l)
: 5Mo MACE free rate = 0.634, se=0.033
38
- 57 -
/
g
group=1:
p 5Mo MACE free rate=0.922 (SE=0.031)
group=2: 5Mo MACE free rate=0.634 (SE=0.033)
- =NORMDIST(6.361,0,1,TRUE)
enter <.999
- P-value = 2*(1-<.999)=<.0001
5Mo MACE free rate SE Z-test
39
6. Event
( )
Variable :
Age(year)
g y
lesion type (1, 2, 3, 4)
Time to MACE(month)
Status(MACE(major j adverse
d cardiac
di events); MACE=1, censored=0)
d
40
- 58 -
2011 ()
41
Analysis
l using Cox proportionall hazards
h d model
d l
(MACE ( ))
:
- MACE
( )
(=Proportional hazard(PH) )
42
- 59 -
/
Result
B)
B al B) 95.0
lesion 36.98
36 98 3.00
3 00 0.000
0 000
lesion1) -2.86 1.01 8.02 1.00 0.000 0.06 0.01 0.1
lesion2) -1.38 0.28 23.8 1.00 0.000 0.25 0.15 0.
lesion3) -0
0.82
82 0 20
0.20 16 9
16.9 1 00
1.00 0 000
0.000 0
0. 0 30
0.30 0 65
0.65
age 0.03 0.01 12.00 1.00 0.000 1.03 1.01 1.05
Lesion 4 MACE .
(P-value = <.001)
llesion(1):
i (1) 1 vs 4 ; P-value
P l = < 0.001
0 001 X 3 = <.001
001
lesion(2): 2 vs 4 ; P-value = < 0.001 X 3 = <.001
lesion(3): 3 vs 4 ; P-value = < 0.001 X 3 = <.001
(P-values were corrected by Bonferronis method due to multiple testing)
43
Hazard ratio(HR)
HR is
i the
h ratio
i off two h
hazard
d rates.
Cf) Hazard
H d rate
t isi the
th instaneous
i t rate
t for
f eventt occurrence. (difficult
(diffi lt
for practical interpretation)
44
- 60 -
2011 ()
Hazard ratio(HR)
= (hazard in lesion 2) / (hazard in lesion 4)
45
N t 2
Note 2: HR P-value
P l
Note 3: P>0.05
P>0 05 95% CI for HR includes HR
HR=1
1.0
0
Note 4: ,
event
( ) lesion
(ex) l i 2 vs 4 ,
lesion 2 event
)
46
- 61 -
/
Cf)
f
RR = the ratio of risks
OR = the ratio of odds (logistic regression)
HR = the ratio of hazard rates (Cox regression)
47
Proportional hazard(PH)
check
(1) Graphical
h l approach
h
- Survival graph
- log (-log
( log (survival)) graph
48
- 62 -
2011 ()
lesion
- PH check: Graphical approach
49
50
- 63 -
/
lesion
- PH check; Statistical test
Schoenfeld
h f ld residual
d l (Harrell and
d Lee(1986))
Step1
partial residual Cox regression
Step2
Event=1 time partial
residual
51
Spearman r
ho
Month ( .000 .000 .011 .251
52
- 64 -
2011 ()
PH check
L i MACE
Lesion
Cox regression
Time-dependent Cox regression analysis
53
_ _ ( ) 0.20
T_COV_*lesion(3) 0.08 6.72 1.00 0.01 1.22 1.05 1.42
54
- 65 -
/
(P-values
(P values were corrected by Bonferroni
Bonferroniss method due
to multiple testing)
55
HR for
f lesion
l i (2 vs 4(=ref))
4( f)) over the
th follow-time
f ll ti
time HR
1 0.089815 2.5
2 0.120032
3 0.160414 2
4 0.214381
5 0.286505 1.5
6 0.382893
7 0.511709 1
8 0.683861
9 0.913931 0.5
10 1.221403
11 1.632316 0
12 2.181472 1 2 3 4 5 6 7 8 9 10 11 12
<-
HR(t) = exp(-2.7
exp( 2 7*(1
(1-0)+0.29
0)+0 29* (1
(1*tt-0
0*t))
t)) = exp(
exp(-2.7+0.29t)
2 7+0 29t)
56
- 66 -
2011 ()
57
58
- 67 -
/
Recurrent events
Outcome events that may occur more than once over the
follow-up time for a given subject.
59
(Competing Risk)
-
. (2 )
- ,
.
60
- 68 -
2011 ()
61
Disease relapse
l competing risk
k transplant
l related
l d
morality (TRM) ?
62
- 69 -
/
Analysis
l using Competing risk
k
63
() KM vs CR
64
- 70 -
2011 ()
1
Site (1=Upper T 2=Mid T 3=Lower T 4=Main) survival curve
65
site=2 2 .
site=2
death
.
->
>
- ,
-,
descriptive
66
- 71 -
/
Summary
G
General
l method
h d S
Survival
i ld data analysis
l i
Summary of data Descriptive statistics Descriptive statistics
mean, variance Survival rate
median, inter quartile Median survival time
range (IQR)
Tests for K independent Parametric method Parametric method
samples t-test Likelihood ratio test(LRT)
One-way ANOVA
p
Nonparametric method
Nonparametric Log-rank test
method
Mann-Whitney test
Kruskal-Wallis test
Regression Multiple linear Cox regression
regression
g Time dependent
p Cox
Logistic regression regression
67
- 72 -
/
2011 ()
SPSS ()
- 73 -
2011 ()
SPSS
SPSS
(
))
OzV~GaGpitGzwzzGzGX`PG
y
- Repeated measures ANOVA (Ex1-1)
- Analysis using Mixed model (Ex1
(Ex1-2)
2)
- Analysis using Generalized estimating
q
equation (GEE)
( ) ((Ex2))
y
- Survival rate estimation & Log-rank test(Ex3-1)
- Cox regression analysis (Ex3-2)
- Time dependent Cox regression analysis
(Ex3-2)
2
- 75 -
/
Example 1-
1 -1
(Repeated measures ANOVA)
:
- ((55
, ) . ( )
- , .
- , .
- 76 -
2011 ()
Independent variable
: age(55 , ), time()
y Method :
- 77 -
/
Method in SPSS
> >
- 78 -
2011 ()
XW
- 79 -
/
XX
N
.00 7
age_1
1 00
1.00 9
- 80 -
2011 ()
y test
->
- ,
(P-value = 0.021)
-
(55 , ) .
(P-value = 0.506)
- 81 -
/
- , (55 , )
(P-value=
. (P value= 0.048)
0 048)
Example 1-
1-2 (Mixed model)
Ex1) (55, >55) 1,
3, 6, 9, 12
X]
- 82 -
2011 ()
Independent variable
: age(55 , ), time()
- 83 -
/
X`
YW
- 84 -
2011 ()
> >
YX
(
( )
)
- 85 -
/
YZ
Y[
- 86 -
2011 ()
Y\
Y]
- 87 -
/
Y^
Model Checking
y Residual plot
: to detect any outliers or a general lack
normality
Y_
- 88 -
2011 ()
Y`
ZW
- 89 -
/
> >
ZX
ZY
- 90 -
2011 ()
ZZ
Z[
- 91 -
/
Example 2 (GEE)
Ex2) 1, 3, 6, 9, 12
:
(55 ,
,)
)
Outcome variable
: 1,
1 3,
3 6,
6 9,
9 12
Independent
depe de t variable
a ab e
: (55 , ), time
- 92 -
2011 ()
Z^
> >
Z_
- 93 -
/
Z`
[W
- 94 -
2011 ()
[X
[Y
- 95 -
/
[Z
96 100 0%
100.0%
0 .0%
96 100 0%
100.0%
id 21
- visit 5
21
2
5
- 96 -
2011 ()
Type III
Survival analysis
- 97 -
/
Example 3-
3-1
Variable :
Group (Multi vessel=1, single vessel=2)
Time to MACE(month)
(MACE ; MACE=1, =
[_
- 98 -
2011 ()
[`
\W
- 99 -
/
\Y
- 100 -
2011 ()
\Z
- 101 -
/
Results
\\
Example 3-
3-2
Variable :
Age(year)
lesion type (1, 2, 3, 4)
Time to MACE(month)
Status(MACE(major adverse cardiac events);
MACE=1,censored=0)
- 102 -
2011 ()
\^
:
- MACE .
- 103 -
/
\`
]W
- 104 -
2011 ()
]X
]Y
- 105 -
/
Lesion 4 MACE .
(P-value = <.001)
][
- 106 -
2011 ()
lesion
- PH check: Graphical approach
> > Cox >
]\
lesion
lesion Survival graph
]]
- 107 -
/
lesion
lesion
log(--log (survival)) graph
log(
]^
lesion
- PH check; Statistical test
Step2
Event=1 time partial
residual
]_
- 108 -
2011 ()
]`
^W
- 109 -
/
> >
^X
^Y
- 110 -
2011 ()
Spearman r
Month ( .000 .000 .011 .251
ho
PH check
y Lesion MACE
Cox regression
- 111 -
/
^\
^]
- 112 -
2011 ()
^^
^_
- 113 -
/
^`
Exp(B) 95.0%
B Wald Exp(B) CI
age 0.03 0.01 10.87 1.00 0.00 1.03 1.01 1.05
lesion 31.57 3.00 0.00
lesion(1) -4.13 2.37 3.04 1.00 0.08 0.02 0.00 1.67
lesion(2) -2.70 0.62 18.81 1.00 0.00 0.07 0.02 0.23
lesion(3) -1.60 0.37 18.39 1.00 0.00 0.20 0.10 0.42
0 27
0.27 0 33
0.33 0 65
0.65 1 00
1.00 0 42
0.42 1 31
1.31 0 68
0.68 2 51
2.51
T_COV_*lesion(1)
80
- 114 -
2011 ()
- 115 -
/
Example 1
1--1
_Z
Example 1
1--2
_[
- 116 -
2011 ()
Example 2
_\
_]
- 117 -
/
Example 3
3--1
_^
__
- 118 -
2011 ()
_`
`W
- 119 -
/
`X
Example 3
3--2
`Y
- 120 -
2011 ()
`Z
`[
- 121 -
/
`\
- 122 -